Eng. Opt., 2002, Vol. 34(5), pp.

461–484

ARCHITECTURAL LAYOUT DESIGN OPTIMIZATION

JEREMY J. MICHALEKa,*, RUCHI CHOUDHARYb and PANOS Y. PAPALAMBROSa
a
Optimal Design Laboratory, Department of Mechanical Engineering, University of Michigan,
Ann Arbor, Michigan 48109-2125, USA; bCollege of Architecture and Urban Planning,
University of Michigan, Ann Arbor, Michigan 48109-2125, USA

(Received 28 August 2001; In final form 26 February 2002)

This article presents an optimization model of the quantifiable aspects of architectural floorplan layout design, and a
companion article presents a method for integrating mathematical optimization and subjective decision making
during conceptual design. The model presented here offers a new approach to floorplan layout optimization that
takes advantage of the efficiency of gradient-based algorithms, where appropriate, and uses evolutionary
algorithms to make discrete decisions and do global search. Automated optimization results are comparable to
other methods in this research area, and the new formulation makes it possible to integrate the power of human
decision-making into the process.

Keywords: Optimization; Architectural design; Floorplan; Layout

1 INTRODUCTION

Spatial configuration is concerned with finding feasible locations and dimensions for a set of
interrelated objects that meet all design requirements and maximize design quality in terms
of design preferences. Spatial configuration is relevant to all physical design problems, so it
is an important area of inquiry. Research on automation of spatial configuration includes
component packing [11–13], route path planning [18], process and facilities layout, VLSI
design ½16; 17
, and architectural layout [3–10]. Architectural layout is particularly interesting
because in addition to common engineering objectives such as cost and performance, archi-
tectural design is especially concerned with aesthetic and usability qualities of a layout,
which are generally more difficult to describe formally. Also, the components in a building
layout (rooms or walls) often do not have pre-defined dimensions, so every component of the
layout is resizable.
Reported attempts to automate the process of layout design started over 35 years ago [3].
Researchers have used several problem representations and solution search techniques to
describe and solve the problem.
One approach to spatial allocation is to define the available space as a set of grid
squares and use an algorithm to allocate each square to a particular room or activity [4–7]
(see Fig. 1). This problem is inherently discrete and multi-modal. Because of the

* Corresponding author. E-mail: michalek@umich.edu

ISSN 0305-215X print; ISSN 1029-0273 online # 2002 Taylor & Francis Ltd
DOI: 10.1080=0305215021000033735
462 J. J. MICHALEK et al.

FIGURE 1 Sample fixed grid allocation layout.

combinatorial complexity, it cannot be solved exhaustively for reasonably-sized layout pro-

blems. Several heuristic strategies have been developed to find solutions without searching
the design space exhaustively. Liggett and Mitchell [4] use a constructive placement strategy
followed by an iterative improvement strategy. In this method, space is allocated for rooms
one at a time based on the best probable design move at each step. Other researchers have
used stochastic algorithms for search [5–7].
Another approach to representing the building layout design space is to decompose
the problem into two parts: topology and geometry. Topology refers to logical relationships
between layout components. Geometry refers to the position and size of each component
in the layout. Topological decisions define constraints for the geometric design space. For
example, a topological decision that ‘‘room i is adjacent to the north wall of room j’’ restricts
the geometric coordinates of room i relative to room j. Researchers have developed decision-
tree-based combinatorial representations and used constraint satisfaction programming
techniques to enumerate solutions without exhaustive search. Baykan and Fox [8] and
Schwarz, Berry, and Saviv [9] developed variations of this model and have been able to enu-
merate solutions for a studio apartment and for a nine-room building respectively. Medjeoub
and Yannou [10] developed a similar model, but they use a technique of first enumerating all
topologies that can produce at least one feasible geometry. The designer is then able to review
the feasible topological possibilities and select those which s=he wants to explore geometri-
cally. This technique reduces computation dramatically, and they have shown success for up
to twenty rooms.
Successful generation of global quality solutions has been achieved for medium-sized
problems; however, there is still a need for a strategy that can handle larger problems
computationally. It would be useful to take advantage of the speed of gradient-based
algorithms on the geometric aspects of the layout, because they involve continuous variables.
This article develops a mathematical model for the geometric decisions in the layout pro-
blem that allows efficient solution with gradient-based and hybrid local–global methods. This
model is then embedded into another model used for topology decisions that is solved with
heuristic global methods. The geometric optimization process allows fast solution of large
complex problems that also enables a true interactive design process described in a sequel
article [2]. The topology optimization component has had limited success due to the
combinatorial nature of the topology decisions. The interactive optimization tool can be
downloaded from http:==ode.engin.umich.edu.
 ARCHITECTURAL DESIGN 463

2 OPTIMIZATION OF GEOMETRY
The geometric optimization problem is posed as a process of finding the best location and
size of a group of interrelated rectangular units. A new decision model is formulated
where all objectives and constraints are continuous functions, and all design variables
have continuous domains.
A Unit is defined as a rectangular, orthogonal space allocated for a specific architectural
function. Examples of architectural functions include living spaces, storage spaces, facilities,
and accessibility spaces. For simplicity, this representation assumes that all Units can be
represented as rectangles or combinations of orthogonal rectangles. This simple representation
can model a large array of architectural layouts, and more complex shapes could be added to
the model to expand this array. Figure 2 shows a Unit represented as a point in space ðx; yÞ, and
the perpendicular distance from that point to each of the four walls: fN ; S; E; and W g. This
model has more variables than necessary to describe the shape; however, it allows an optimiza-
tion algorithm to change the position of a Unit independently without affecting its size (by
changing x or y), and it can change any of the four wall positions independently (by changing
N ; S; E, or W ). Although this model increases the problem dimensionality, it offers a lot of
flexibility to make the best design moves at each step of the optimization.
Units are grouped into several categories based on their function: Rooms, Boundaries,
Hallways, and Accessways. Rooms are Units used for sustained living activity as determined
by the designer. The differentiation between living space vs. non-living space is important
only in optimization objectives that maximize the amount of space used for living relative
to all other space. A Boundary is a Unit that has other Units constrained inside of it, and
it is not considered living space. A Hallway is a Unit with no physical walls that is not a
living space. Hallways function as pathways. An Accessway is a Hallway that is constrained
to geometrically intersect two Units. Accessways are generally restricted to be small,
and they are forced to intersect two other Units. They function to keep the two Units adjacent
and connected, and to ensure that there is room for a door or opening between the rooms.
In Figure 3, the external rectangle represents the building Boundary, the living room, bed-
room, and bathroom are Rooms, the hall is a Hallway, and the three Units labeled ‘‘A’’ are
Accessways that define space for a doorway between Units.
Units that are along external walls may also have windows for natural lighting. Window
height can be fixed for each Unit, and window width is a variable. oN ; oS ; oE ; oW represent
the width of the north, south, east and west windows, respectively.

FIGURE 2 Representation of a Unit.

464 J. J. MICHALEK et al.

FIGURE 3 An example layout showing four different types of Units.

2.1 Mathematical Geometry Optimization Model

The design optimization problem is formulated as
minimize f ðxÞ
subject to hðxÞ ¼ 0
ð1Þ
gðxÞ  0
x 2 <n

where x is the vector of design variables, n is the number of variables, and h(x) and g(x) are
vectors of equality and inequality constraints.

2.1.1 Design Variables

Variables for each Unit include a reference point location ðx; yÞ, distances to each wall
ðN ; S; E; W Þ, and the size of any windows added to each Unit:
n
x ¼ [ fxi ; yi ; Ni ; Si ; Ei ; Wi ; oNi ; oSi ; oEi ; oWi g
i¼1 ð2Þ
xi ; yi ; Ni ; Si ; Ei ; Wi 2 <; oNi ; oSi ; oEi ; oWi 2 <þ

The window variables drop out when the window is not physically present for a specific Unit
and direction. In order to simplify calculations and notation, several ‘‘intermediate’’ variables
are used to describe geometry that results from the design variables. The following resultant
variables are calculated from the design variables.

yNi ¼ yi þ Ni Unit north wall location ð3Þ

y S i ¼ y i  Si Unit south wall location ð4Þ
xEi ¼ xi þ Ei Unit east wall location ð5Þ
xWi ¼ xi  Wi Unit west wall location ð6Þ
li ¼ Wi þ Ei Unit length ð7Þ
wi ¼ Ni þ Si Unit width ð8Þ
 ARCHITECTURAL DESIGN 465

These relations are linear, so linear functions of these intermediate variables are also linear
functions of the original variables.

2.1.2 Geometric Design Constraints

The following constraint groups form a toolbox of constraints that can be applied where
appropriate for a particular layout problem. Using the software described in the companion
paper [2], default constraints are automatically added to the model whenever the designer
adds a room, window, etc. The designer can also add, delete, or modify constraints individually.
The Force Inside Constraint Group forces Units into the main building Boundary or other
grouping Boundaries. In order to force Unit i inside Unit j, the following four constraints
must all be satisfied:

yNi  yNj ; ySj  ySi ; xEi  xEj ; and xWj  xWi ð9Þ

The Prohibit Intersection Constraint functions to prevent two Units from occupying the
same space. By default, one Prohibit Intersection Constraint is added for each combination
of Rooms, Hallways, and Accessways, except where two Units are forced to intersect, or
where one Unit is forced inside of another. In order to prevent Unit i from intersecting
Unit j, at least one of the following constraints must be satisfied

ðxWi  xEj Þ OR ðxWj  xEi Þ OR ðySi  yNj Þ OR ðySj  yNi Þ ð10Þ

The logical disjunction can be represented in negative null form using a min function

minðxEj  xWi ; xEi  xWj ; yNj  ySi ; yNi  ySj Þ  0 ð11Þ

This nonlinear, non-smooth formulation is undesirable for gradient-based calculations;

however, the nature of the constraint makes it unavoidable. With this formulation, the con-
straint function acts as a smooth linear function except when the close corners of two
Units are nearly diagonal.
The Force Intersection Constraint Group is used when Units are forced to intersect in
order to ensure access (as Accessways do), or to make a more complex geometric shape
by combining rectangular Units. Forcing intersection is the opposite of prohibiting intersec-
tion, so forcing intersection can be written as the conjunction of the following constraints

ySi  yNj ; ySj  yNi ; xWi  xEj ; and xWj  xEi ð12Þ

Although these constraints ensure intersection of the two Units, they permit intersection at a
point. Designers of architectural spaces are generally interested in intersection that provides
enough room for a doorway or opening. To model this, an additional constraint is included to
enforce overlap in one of the Cartesian directions that is at least as large as the doorway or
opening. Therefore, in addition to intersection, at least one of the following conditions must
be satisfied

yNj  ySi  maxðdi ; dj Þ Unit i overlaps north wall of Unit j; ð13Þ

yNi  ySj  maxðdi ; dj Þ Unit i overlaps south wall of Unit j; ð14Þ
xEj  xWi  maxðdi ; dj Þ Unit i overlaps east wall of Unit j; ð15Þ
xEi  xWj  maxðdi ; dj Þ Unit i overlaps west wall of Unit j: ð16Þ
466 J. J. MICHALEK et al.

where di is the minimum size for a door or opening in Unit i. This disjunctive set of
constraints can be represented in negative null form using a min function similar to Eq. (11).

minfmaxðdi ; dj Þ  xEj þ xWi ; maxðdi ; dj Þ  xEi þ xWj ;

maxðdi ; dj Þ  yNj þ ySi ; maxðdi ; dj Þ  yNi þ ySj g  0 ð17Þ

Although this constraint function is nonlinear and non-smooth in part of the design space, it
is linear in most of the design space (similar to Eq. (11)). The complete Force Minimum
Intersection Constraint Group is represented as a set of constraints that force intersection
(Eq. (12)) and another constraint to ensure that the overlap is large enough for access
(Eq. (17)).
The Force To Edge Constraints are used to force a Unit to the edge of a Boundary because
of a window or external door. It is assumed that the first Unit i has already been forced inside
Unit j by another constraint. In order to force a Unit to a particular wall, one of the following
constraints can be added as appropriate.

yNi ¼ yNj ; ySi ¼ ySj ; xEi ¼ xEj ; or xWi ¼ xWj ð18Þ

If connection to an edge is important, but the specific edge is not important, (for instance, a
building may require an external door, but it is not important which direction the door faces),
then the following constraint can be added to represent the disjunction in Eq. (18).


min ðxEi  xEj Þ2 ; ðxWi  xWj Þ2 ; ðySi  ySj Þ2 ; ðyNi  yNj Þ2 ¼ 0 ð19Þ

This representation is non-smooth at Unit corners; however, it is quadratic in most of the

design space.
The Bound Size Constraint Group includes three kinds of constraints to bound the area of a
Unit: minimum area, minimum length=width, and maximum length=width. It is assumed that
a maximum area constraint would not be used to bound the area. Instead, Unit area is only
reduced to improve objective functions, such as cost objectives. Minimum area, Amin ,
minimum length=width, lmin , and maximum length=width, lmax can be set for each Unit.

Amini  li wi  0 minimum area ð20Þ

lmini  li  0 and lmini  wi  0 minimum length=width ð21Þ
li  lmaxi  0 and wi  lmaxi  0 maximum length=width ð22Þ

The Minimum Ratio Constraint Group can be used to maintain a desired aesthetic scheme
or prevent long, narrow Rooms that may not be usable. The Minimum Ratio Constraint
Group consists of two constraints.

Rmini li  wi  0 and Rmini wi  li  0 ð23Þ

The Build Cost Constraint is used to keep the construction cost below some value, Gbudget .
For simplicity, build cost is measured only in terms of material cost. Material costs for walls
kwall and for windows ko are specified as dollars per square foot of material, and other costs
are ignored. The build cost constraint is calculated as

kwall ðAN þ AS þ AE þ AW Þ þ ko ðAoN þ AoS þ AoE þ AoW Þ  Gbudget ð24Þ

 ARCHITECTURAL DESIGN 467

where AN ’ AS ’ AE’ AW are the areas of the external walls in each compass direction and
Ao Ao Ao AoW are the areas of windows facing each compass direction. These quantities
N’ S’ E’
are computed in Eq. (30) and Eq. (31).
The Feasible Window Constraints ensure that the window width cannot be larger than the
wall it is on. In addition to the simple bound restricting window size to be positive, this
ensures feasible window size. Each window added to a Unit is given one of the following
Feasible Window Constraints as appropriate.
oNi  li ; oSi  li ; oEi  wi ; or oWi  wi ð25Þ

The Bound Lighting Constraint is used to ensure minimum natural lighting for specific
rooms. A simple estimation of the amount of daylight entering a Unit with windows is cal-
culated using environmental and material information. The following procedure is used:
First, available daylight at the window exterior is determined. IESNA [31] provides three
standard skies for use in the evaluation of daylight designs. Approximate available daylight
can be determined from these based on altitude and azimuth angles.

Evkdm ¼ vertical sky illuminance (direct)

Evksm ¼ vertical sky illuminance (sky)

for month m. The coefficient of utilization, CU , a function of the room geometry and window
size, determines the fraction of the available daylight that enters the room. CU can be found
in pre-tabulated data [30] based on room depth, window width, and window height. The net
transmittance for a window facing direction j is calculated as

0:9mG Aoj
mj ¼ ð26Þ
Aj

where j takes on each of the directions fN ; S; E; W g; mG is the transmittance of the window

(material property), Aoj is the area of the glass in direction j, and Aj is the area of the wall in
direction j. Finally, the daylight at the room center is calculated. The horizontal illuminance
at the center of room i is calculated as
!
XX
2
Ei ¼ Aoj ðEvkdmi þ Evksmi Þmj CU 10 ð27Þ
j m

for room i, where j takes on each direction fN ; S; E; W g, and m spans the 12 months of the
year. The illuminance is then converted into watts, Yi :

Ei 103
Yi ¼ ð28Þ
Ai beff

where Ai is the area of room i, and beff is the efficacy of the light source (assumed to be 80).
The required natural lighting per square foot, yreqi , is defined for each Unit by the designer
(default 1 Watt=sq.ft.). Assuming uniform light distribution, total required natural lighting
can be calculated as Ai yreqi . The minimum percentage of required lighting that is provided
by natural light, jmini , can be specified by the designer. The final constraint is written as:

Yi
 jmini ð29Þ
Ai yreqi
468 J. J. MICHALEK et al.

2.1.3 Geometric Design Objectives

Several objectives have been defined that can be used independently or together depending
on the designer’s goals.
The Minimize Heating Cost Objective estimates heating loss during cold months. The
annual energy cost to heat the building is calculated as a function of the building
Boundary Unit shape, volume, surface area, and material as well as environmental condi-
tions. Simplified calculations (ASHRAE [30]) are used as an approximation. The procedure
for calculating heating loads is as follows. It is assumed that windows on all Units are
constrained against external walls, so the net area of windows on each external wall is:
X
AoD ¼ oDi hoi ð30Þ
i2UnitD

Here D takes on the four directions fnorth, south, east, westg, and UnitD refers to Units that
have windows in direction D. The net area of each external wall is

AD ¼ l1 h1  AoD ð31Þ

Here D takes on the directions fnorth, south, east, westg, and 1 indicates Unit 1, which is
assumed to be the building Boundary Unit. The heat loss calculation assumes that all heat
is lost from the external walls and windows (no heat is lost through the roof ). This model
could be changed depending on what type of building is being modeled. The coefficient
of transmittance for the wall, Uwall , and window, Uo , are tabulated based on the materials
used. The annual heat loss is
X
Qheat ¼ DTi ððAN þ AS þ AE þ AW ÞUwall þ ðAoN þ AoS þ AoE þ AoW ÞUo Þ ð32Þ
i

where i is the set of months where heat is used, and DTi is the average internal=external tem-
perature difference for month i. Finally, the cost to maintain temperature is calculated. Gas
heat is assumed, and the cost of gas per cubic foot, kgas and efficiency of the heater in
Watts per cubic foot of gas, Zheater , can be specified. The heating cost objective function is then

kgas Qheat
minimize Gheat ¼ ð33Þ
Zheater

The Minimize Cooling Cost Objective estimates heat gain during hot months. The proce-
dure for calculating cooling loads is more complicated than heating loads because heat due to
solar gain must be taken into account. The procedure works as follows. First, the net area of
windows on each external wall is calculated using Eq. (30), and the net area of each external
wall is calculated using Eq. (31). Next, the solar heat gain through the windows is estimated.
Several parameters are important in calculating solar heat gain. Depending on the orientation
of the windows (N, S, E, or W), the Solar Heat Gain Factor, bshgf , can be found in tables for a
given location [30]. The shading coefficient, bsc , is a property of the glass [30], and the time-
lag factor, btlf , is a tabulated function of glass type and window orientation [30]. The annual
solar heat gain, Qsolar , is calculated as
!
X
Qsolar ¼ bsc ðAoN bshgf N btlf N þ AoS bshgf S btlf S þ AoE bshgf E btlf E þ AoW bshgf W btlf W Þ ð34Þ
i
 ARCHITECTURAL DESIGN 469

where i is the set of months where air conditioning is used. Next, the conductive heat gain
through the building exterior is estimated. The orientation of each exterior wall and windows
is accounted for in the factor. The cooling load due to conduction is calculated as
X
Qcond ¼ DTi ðUo ðAoN btlf N þ AoS btlf S þ AoE btlf E þ AoW btlf W Þ
i
þ Uwall ðAN btlf N þ AS btlf S þ AE btlf E þ AW btlf W ÞÞ ð35Þ

where i is the set of months where air conditioning is used. Finally, the cost to maintain room
temperature is calculated. Electric cooling is assumed, and the rate of electricity, kelec , and
efficiency of the air conditioning unit, Zac , can be specified. The cooling cost objective
function is then
kelec ðQsolar þ Qcond Þ
minimize Gcool ¼
Zac

The Minimize Lighting Cost Objective minimizes the cost spent on lighting the building by
encouraging natural lighting. The amount of natural lighting in room i, Yi , is calculated as in
Eq. (28). The minimum daylight requirement per square foot, blight , is set by the designer
based on usage intention. The total required cost if all of this light is provided by electric
lighting can be calculated as:
!
X
Gelec blighti Ai bH 103 ð36Þ
i

where i is the set of Units, and bH is the number of hours of use per month. The total cost is
then the maximum possible electricity cost minus the cost savings from natural lighting:
!
X
minimize Glight ¼ Gelec  Yi bA 103 ð37Þ
i

where i is the set of Units, and bA is the number of hours of available light per month.
The Minimize Wasted Space Objective minimizes building space that is not living space.
This could be space used for hallways or un-allocated space inside the building Boundary.
Wasted space is calculated as the area of the building Boundary minus the total area used
as living space. The objective is formulated as
!
X
minimize l1 w1  li wi ð38Þ
i 2 Rooms

where 1 indicates Unit 1, which is assumed to be the building Boundary Unit.

The Minimize Accessways Objective brings connected Units together. Accessways may be
constrained to be small to keep Units together. Alternatively, the Minimize Accessway
Objective can be used to bring Units together if possible, but allow them to be separated
if necessary, providing that there is an Accessway between them. This method allows
Accessways to function similarly to Hallways depending on the design situation. The objec-
tive is formulated as
!
X
minimize li wi ð39Þ
i 2 Accessways
470 J. J. MICHALEK et al.

The Minimize Hallway Objective is used to provide extra living space where possible. The
objective is formulated as
!
X
minimize li wi ð40Þ
i 2 Hallways

Multiple objectives can be selected and combined into a single objective function using a
weighted sum of the individual objective functions.

X
N
f ðxÞ ¼ wj fj ðxÞ ð41Þ
j¼1

where fj ðxÞ is the jth objective function, wj is the weighting (relative importance) of the jth
objective function, and N is the total number of objective functions. Appropriate weights may
be difficult to set for objective functions measured in different units. After obtaining results,
weights can be adjusted to compensate and to guide the design to desired results. The objec-
tives presented here do not compete in most of the design space, except for cost objectives,
which are all measured in dollars. This makes multi-objective optimization much easier. In
practice weights only need to be adjusted to keep the function values in the same order of
magnitude to avoid computational problems.

2.2 Geometry Model Solution Methods

2.2.1 Local Optimization Method
CFSQP, a C implementation of Feasible Sequential Quadratic Programming [25], was used to
solve the building geometric layout problem presented above. FSQP is similar to SQP except
that once a feasible design is found, search directions are altered to maintain feasibility at
every iteration. If the initial design is infeasible, a penalty function strategy is used to find
a feasible design. In addition, CFSQP also handles linear constraints separately so that
they are solved more efficiently. A sample optimization of a particular layout problem is
shown in Figure 4.
CFSQP is very fast for moderately sized problems using this formula, and it is relatively
stable; however, sometimes the algorithm becomes stuck, partly due to non-smooth con-
straints (Eqs. (11), (17), (19)). Still, in practice the algorithm almost always converges
quickly, and convergence problems can usually be avoided by perturbing the design slightly
to move it away from non-smooth areas of the design space.
Gradient-based search algorithms find locally optimal designs. This means that the design
is better than any neighboring design; however, the solution is highly dependent on the start-
ing point, and there is no guarantee that the design is of global quality. The design space of
this problem contains many local optima, some of which have poor global quality. Also, if
the starting point is highly infeasible, then the algorithms often cannot find feasible designs.

2.2.2 Global Optimization Methods

Global optimization methods have been developed to overcome the limitations of local
search and to find solutions of global quality. Several global search strategies were used to
generate geometric layouts.
Both Simulated Annealing (SA) and Genetic Algorithms (GA) were implemented to
search the geometry design space for global solutions. Because of the highly constrained
 ARCHITECTURAL DESIGN 471

FIGURE 4 Progression of the CFSQP algorithm optimizing a sample apartment complex building to minimize
annual cost and wasted space; (a) shows the initial layout sketch provided by the designer (accessways shown as lines
between Units); (b) is an intermediate feasible iteration (accessways shown also as rectangles); and (c) shows the
completed design (accessways shown as wall openings for clarity).
472 J. J. MICHALEK et al.

FIGURE 4 (Continued)

nature of the formulation, neither algorithm was successful at finding feasible designs for
small problems. This does not mean that the algorithms cannot be successful at laying out
architectural spaces; however, both algorithm are ill-suited to the formulation presented here.
A hybrid SA=SQP search strategy was developed to take advantage of the global qualities
of SA and the efficiency of SQP in order to generate local optima of global quality. The
method works as shown in Figure 5.
In this method, SA is used to search for a good starting point, and SQP is used to find the
local minimum near each starting point. In this way SA can search the space more globally
with large moves while SQP worries about the details. A sample objective function is shown
in Figure 6. In this example, SQP can find six different local optima depending on where the
starting point is chosen. Each point that SA selects is evaluated by locally optimizing it, so
SA observes any point in the vicinity of a local optimum to have the objective value of that

FIGURE 5 Description of the SA=SQP hybrid algorithm.

 ARCHITECTURAL DESIGN 473

FIGURE 6 Hybrid SA=SQP sample function with multiple local minima.

local optimum. In a sense, the objective function is being screened for SA. Notice in the
example that the function SA observes has only two local optima instead of six. Also, an
algorithm searching the resultant function can make larger design moves without as much
danger of overstepping important features.
The hybrid SA=SQP method generated local optima of reasonable global quality for up
to seven room apartment layouts (70 variables, 269 constraints – see Ref. [1] for resulting
layouts). It is important to understand that these designs were generated automatically with
no feasible initial starting point. This is a substantial improvement. Using SA alone, we
were unable to produce even a feasible design. SQP is quick at generating solutions; how-
ever, the designer must define where Rooms should be placed relative to one another.
In this problem, the arrangement is not specified by the designer. The algorithm is
able to automatically generate a quality feasible arrangement and optimize that geometry
locally.
Another way to search for solutions of global quality is to use a variation of an optimiza-
tion technique referred to as the Maximum Distance Distribution Method (MDDM ½23; 24
).
This method was developed for discrete problems, but it also works for continuous problems.
The concept is to use a local optimization algorithm to find a local minimum x using the
formulation in Eq. (1). Once the local minimum is found, a new optimization problem is for-
mulated to maximize the distance from x subject to an extra constraint that the new point
must have an objective value at least as good as f ðx ).

maximize ðx  x Þ2
subject to hðxÞ ¼ 0;
gðxÞ  0; ð42Þ
f ðxÞ  f ðx Þ  0
x 2 <n

If optimizing Eq. (42) yields a solution xy in a new area of the design space, then optimizing
Eq. (1) again with xy as a starting point will often yield a better local minimum. This process
can be repeated by iteratively solving Eqs. (1) and (42) to obtain better solutions. MDDM is
not guaranteed to converge to the global optimum; however, in practice there are many
situations where this method is successful at improving the quality of the local optimum
returned. An example is provided in Figure 7. The method is especially useful if f ðx Þ is
flat in some feasible direction at x .
474 J. J. MICHALEK et al.

FIGURE 7 Demonstration of the MDDM method for finding improved local optima. An initial design (a) was
optimized using CFSQP. The result is a local optimum (b) (the design cannot be improved by small changes in the
design variables). The MDDM method was used to generate design (c), an improved local optimum for this example.

Another design exploration program was written to produce design alternatives by searching
the space using a strategy of random design changes. The program makes design moves of
three types: (1) swap the positions of two Units, (2) perturb the position of a Unit, and (3)
reduce the size of a Unit. After each design move, the program attempts to re-optimize
using the geometric optimization algorithm. The algorithm first attempts to find a feasible
design using penalty methods. If it is unable to find a feasible design, the program makes
one of the three design moves at random. When a feasible design is found, it is saved, and
the program continues by making more random design moves. This strategy was used to
generate designs for a simple three-bedroom apartment layout. The program generated 200
design alternatives overnight. Although this strategy is not rigorous, it is a useful tool for
generating a spread of design alternatives that can be explored further with the interactive
design tool (see Ref. [2]).

3 OPTIMIZATION OF TOPOLOGY

The topology optimization problem is presented as a process of finding the best set of rela-
tionships between rooms in a space. In this formulation, relationships include connectivity,
and initial rough location. Connectivity defines which rooms are directly connected by a
doorway or open pathway. Rough location defines rough arrangement of rooms. Other mod-
els (½9; 10
) have used decision variables to define topological spatial relationships (i.e. adj-
to-north-of, adj-to-south-of, etc:). However, the use of rough room position to describe spa-
tial relationships does not enforce these relationships during geometric optimization, so the
geometric optimization algorithm has more freedom to manipulate the geometry.
Topologies could be evaluated based on topological qualities, such as openness, proximity,
directionality, or symmetry; however, even though these aspects are often thought of as
topological, they are difficult to evaluate without rough geometry. It is best to evaluate
objectives using a geometric layout, therefore each topology is evaluated based on the
 ARCHITECTURAL DESIGN 475

best geometry that can be generated from it. Using this method, layouts can be optimized for
any objective that can be formulated in terms of geometry or topology.
Figure 8 shows the topology optimization process. A discrete optimization algorithm uses
information from previous topologies to generate new topologies. Each new feasible topo-
logy, X, is translated into a geometric optimization problem. A locally optimal geometry
x is found using CFSQP, and the quality of that geometry fg ðx Þ defines the quality
of the topology that generated it, ft ðX Þ. The discrete optimization algorithm searches for
the topology that generates the best geometry.

3.1 Mathematical Topology Optimization Model

3.1.1 Variables
The variables for the topology optimization problem are the initial grid position of each
room, and the connectivity between each room and every other room=external wall.

xi ; yi 2 Zþ
fij 2 f0; 1g ð43Þ
8 i 2 froomsg; 8 j 2 ðfrooms > ig [ fextwallsgÞ

where ðxi ; yi Þ represents integer Cartesian coordinates of room i, and fij represents the
existence of a connection between room i and room j (or external wall j). Figure 9

FIGURE 8 Building topology optimization method.

476 J. J. MICHALEK et al.

FIGURE 9 A 4-room example showing design variables in the topology formulation. (a) Room position grid
showing (x, y) for each room. Phantom lines show room connections. The dashed line shows the implied boundary.

shows a visual representation of the design variables. It is important to note that topological
decisions about relative positions of rooms (i.e., room i is-north-of room j) are represented
here using absolute positions of rooms. Several other methods of representing topological
decisions (½9; 10
) do not use absolute positions; however, it is necessary in this strategy
because the geometric optimization algorithm requires a starting design with geometric
information. The use of absolute positions has several consequences: (1) The mapping
from topology to geometry is not injective (one-to-one). It is possible for more than
one topology to generate the same geometry. This means that computation time can be
wasted searching similar topologies. (2) The mapping from topology to geometry is not
surjective (onto). Because each room is represented as a grid point, each topology could
be interpreted geometrically in several ways. It is not clear, however, that every possible
 ARCHITECTURAL DESIGN 477

geometric alternative can be generated using the topology definition in this paper. (3) The
space of topology combinations is exponential.
Because of these limitations, this representation is not well suited to problems where all
solutions need to be enumerated. It is not clear that the representation can enumerate all pos-
sible topology alternatives; however, this method is powerful for larger problems where heur-
istic search is necessary. This is because in practice heuristic search algorithms can often find
reasonable quality designs quickly, while enumeration algorithms must systematically
explore designs one by one, which can often take too long to be practical.

3.1.2 Topology Design Constraints

The following constraints form a toolbox of constraints that can be applied where appropriate
for a particlar topology layout problem.
The Overlap Constraint ensures that no two rooms occupy the same space.

jxi  xj j þ jyi  yj j  1 8 i 6¼ j ð44Þ

Connectivity Constraints are defined by the designer for each problem. The constraints
describe how a certain room is required to be connected to an outside wall, to another
room, or how certain rooms are required to not be connected. For example,

fij ¼ 1 room i required to connect to room j ð45Þ

fij ¼ 0 room i required not to connect to room j ð46Þ

fiN þ fiS þ fiE þ fiW  1 room i required to connect to at least one external wall ð47Þ

Path Constraints are defined by the designer for each problem. A path may be required
between all combinations of rooms, or a path may be required between certain rooms. For
example, a path could be required from the bedroom to the kitchen without passing through
a bathroom or closet. These constraints involve room connectivity, and they are generated for
each specific constraint with an algorithm (see Ref. [1]).
Planarity Constraints ensure that the geometry can be realized with a two-dimensional
(planar) floorplan. One way to ensure planar feasibility is to draw lines between connected
nodes on the position grid and ensure that no two lines cross. These lines will be allowed
to share endpoints as long as they do not share any interior point. This constraint is difficult
to represent with a closed form mathematical function. (see Ref. [1])
Envelope Constraints ensure that Units that are forced to be connected to an external wall
must lie on the external envelope of Units on that wall. The four constraints below are added
for each unit i

fiN ðmaxð y1 ; y2 ; . . . ; yn Þ  yi Þ ¼ 0 ð48Þ

fiS ðyi  minð y1 ; y2 ; . . . ; yn ÞÞ ¼ 0 ð49Þ

fiE ðmaxðx1 ; x2 ; . . . ; xn Þ  yi Þ ¼ 0 ð50Þ

fiW ðxi  minðx1 ; x2 ; . . . ; xn ÞÞ ¼ 0 ð51Þ

478 J. J. MICHALEK et al.

3.1.3 Objective
The objective of the topology optimization problem is to minimize the objective value of the
resultant local optimal geometry formed by the topology.

minimizeð fg ðSQPðX ÞÞÞ ð52Þ

where fg is the objective value of the geometry, and SQP is a function that returns the local
optimum geometry for the topology X. Notice that x and y determine starting locations for
rooms in the geometry formulation while f defines constraints for the geometry as well as win-
dows and accessways (see Fig. 10). The optimization objective can be anything defined by the
geometry. Typically, this research has optimized for the topology that produces the most cost
efficient layout. Only feasible topologies are passed to the geometric optimizer (SQP). If the
topology violates any constraints, then the design is evaluated using penalty functions.

3.1.4 Penalty Functions

In this formulation, penalty functions are used to represent constraints, and infeasible designs
are not passed to the geometric optimizer. Figure 11 shows how this penalty function works.
Here the topology objective function ft ðX Þ is maximized. The following procedure is used to
evaluate a topology. If the design is infeasible, ft returns a negative value that is penalized for
each constraint violated, and for the extent of violation. If the design is feasible, X is passed
to the geometric optimization algorithm. Assuming the geometric algorithm finds a feasible
geometry ðx Þ; ft returns a bonus value (B) minus the objective function value of the geo-
metric optimum fg ðx Þ. The bonus value is set so that it is larger than any objective value
fg ðxÞ. Using this method, all infeasible topologies return negative function values, all feasible

FIGURE 10 Schematic showing the relationship between the topology and geometry optimization algorithms.
 ARCHITECTURAL DESIGN 479

FIGURE 11 Formulation of the topology objective function.

topologies return positive objective function values, and feasible topologies that result in bet-
ter geometries (lower fg ðxÞ) have a better objective function value (higher ft ðX Þ).

3.2 Topology Model Solution Method

The discrete topology design space is combinatorial, multi-modal, and highly constrained, so
it must be searched with a global scope. The space of topologies could be searched exhaus-
tively with a constraint satisfaction programming enumeration algorithm [27] or branch and
bound; however, combinatorial explosion will cripple the algorithm for problems of signifi-
cant size. Furthermore, enumeration is unnecessary in a problem where many of the implicit
design goals (such as aesthetic intent) are not generally defined mathematically, but instead
must be judged. It is not meaningful to produce a strict global optimum; instead, it is more
useful to produce an array of quality design alternatives to explore. For this reason, evolu-
tionary algorithms were selected. Evolutionary algorithms search heuristically, and they
can be stopped at any point during the optimization process to return a population of best
designs found. This heuristic search, combined with penalty functions, can often find quality
feasible designs to large problems that are intractable for systematic search methods.
An evolutionary algorithm for topology layout was implemented using the GAlib optimi-
zation package [28]. A SteadyStateGA was selected. A Roulette Wheel selector was used to
select high quality designs with greater probability than low quality designs. When sexual
crossover is used (randomly), two parents are selected from the population, and two new chil-
dren are produced using mixed room connectivity from both parents. When asexual crossover
is used (randomly), one parent is selected from the population, and one new child is produced
by swapping connectivity values between rooms or by swapping room positions. After cross-
over, new designs are mutated slightly. Room locations (x,y) are incremented or connecti-
vities are flipped with low probability.
The evolutionary algorithm implementation is able to generate quality feasible designs for
medium-sized problems.
480 J. J. MICHALEK et al.

TABLE I Room Specifications for Demonstration Problem.

Min. area Min. length and width Max. length and

Apt Room (sq. ft.) ( ft.) width ( ft.)

– Public Entry 9 3 100

1 Living Room 160 12 40
1 Dining Room 100 10 30
1 Kitchen 100 8 40
1 Bedroom 120 10 40
1 Bathroom 30 5 20
2 Living Room 160 12 40
2 Dining Room 100 10 30
2 Kitchen 100 8 40
2 Bedroom 1 120 10 40
2 Bedroom 2 120 10 40
2 Bathroom 30 5 20
3 Living Room 160 12 40
3 Dining Room 100 10 30
3 Kitchen 100 8 40
3 Bedroom 1 120 10 40
3 Bedroom 2 120 10 40
3 Bathroom 30 5 20

4 DEMONSTRATION EXAMPLE

A realistic problem was implemented to test the scalability of the automated building genera-
tion algorithm. The example involves a small apartment complex with three separate apart-
ments. Rooms and specifications are shown in Table I. Constraints that are specific to this
problem are listed in Table II. This problem was run for 20,000 generations (100 designs
each generation) to search for global solutions. Feasible designs take much longer to evaluate
than infeasible designs (because feasible designs are passed to the geometric optimization
algorithm), so a second termination criterion was added to terminate after 50 feasible designs
were found. This criterion was intended to make search time more consistent between runs.
The sample topology and resulting geometry solution shown in Figure 12 were generated
using the automated design tool.

TABLE II Topology Specifications for Demonstration Problem.

Constraint type Constraint

Overlap No two Units can occupy the same space

Connectivity Public Entry must connect to the Living Room of each apartment
Connectivity Public Entry must connect to an external wall
Connectivity All bedrooms must connect to an external wall
Path In each apartment, there must be a path from the Kitchen to the
Living Room that may pass through the Dining Room
Path In each apartment, there must be a path from the Bathroom to the
Living Room that may pass through the Dining Room and Kitchen
Path In each apartment, there must be a path from the Dining Room to the
Living Room that may pass through the Kitchen
Path In each apartment, there must be a path from each Bedroom to the
Living Room that may pass through the Dining Room
Accessways Accessway lines connecting Units cannot intersect
Envelope Units that are connected to an external wall must lie on the boundary
envelope of rooms
 ARCHITECTURAL DESIGN 481

The algorithm generates local optimal solutions, but the global search is quite limited due
to combinatorial complexity. Once a feasible topology is found, it has a much higher prob-
ability of being selected as a parent design by the evolutionary algorithm because it has a
much higher fitness value than infeasible designs. Thus, new designs tend to be very similar
to the first feasible design found, and other designs are usually discarded. The result is that
the algorithm tends to fixate on the first feasible solution it finds, exploring mostly variations
of that solution. The algorithm can be run several times to produce design alternatives, but
generally when it is run once, the final population converges to variations of one main design
theme. This is a serious limitation for global search, and the algorithm is more useful as a
feasible-design-finder than as a true optimizer. For smaller problems, the evolutionary algo-
rithm is still able to search a significant range of the design space to find global quality solu-
tions. For this problem, the evolutionary algorithm can consistently find solutions in under
20,000 generations (2  106 design evaluations).
The geometric optimization problem does not have the same combinatorial nature that the
topology problem has, and it is able to handle much larger problems. The example shown
in Figure 4, contains 23 rooms, three hallways, one boundary, and 25 accessways for a
total of 52 units. This geometric optimization problem contains 312 variables and 1578 con-
straints.

FIGURE 12 Sample design topology and final geometry generated by the automated design tool.
482 J. J. MICHALEK et al.

FIGURE 12 (Continued)

5 CONCLUSIONS

Two automated optimization algorithms have been used to automate the generation of design
layouts: the geometry and topology algorithms. The geometry algorithm, built on rigorous
gradient-based algorithms, is efficient and robust, and has been successful at optimizing geo-
metry for large problems. In its present state, it is most useful as an aid for design explora-
tion, rather than design automation, because results are highly dependent on the starting point
defined by the designer. Several tools have been implemented for searching the geometric
space more globally, including a hybrid SA=SQP algorithm and a strategic programming
method. These tools have been successful at automatically finding alternative arrangements
for rooms and exploring many local minima.
A second topology optimization algorithm was built on top of the geometric algorithm to
search feasible topology alternatives and find the feasible topology that generates the best
geometry. The results are interesting, but limited. One advantage to the approach presented
here is that the final design generated by the algorithm can be used as a starting point for
interactive design exploration (see Ref. [2]).
Possible improvements include the addition of new shapes, objectives, and constraints to
the design toolbox to address more complex geometry, material selection, building codes,
structural elements, routing of wires, pipes and ducts, and more accurate models.
Additionally, the topology search can be improved if the topology can be defined in a new
 ARCHITECTURAL DESIGN 483

way so that topology decisions create linear constraints in the geometry optimization pro-
blem. Using a different definition of topology, such as in Refs. [3–5], it is possible to elim-
inate the rough position topology variables, and ensure that the mapping from topology to
local optimal geometry is both injective and surjective (meaning that each valid topology
will create a different locally optimal geometries, and that all possible local optimal geometry
could be created by a specific topology).

Acknowledgements
Special thanks to Professor Kazuhiro Saitou, John Whitehead, Panayiotis Georgiopoulos, and
Adam Cooper for their contributions to the topology optimization model and solution
strategy.

References
[1] Michalek, J. (2001). Interactive layout design optimization. MS thesis, University of Michigan.
[2] Michalek, J. and Papalambros, P. Y. (2002). Interactive design optimization of architectural layouts. Engineering
Optimization (this issue).
[3] Levin, P. H. (1964). Use of graphs to decide the optimum layout of buildings. Architect, 14, 809–815.
[4] Liggett, R. S. and Mitchell, W. J. (1981). Optimal space planning in practice. Computer-Aided Design, 13(5),
277–288.
[5] Sharpe, R., Marksjo, B. S., Mitchell, J. R. and Crawford, J. R. (1985). An interactive model for the layout of
buildings. Applied Mathematical Modeling, 9, 207–214.
[6] Jo, J. H. and Gero, J. S. (1998). Space layout planning using an evolutionary approach. Artificial Intelligence
in Engineering, 12, 149–162.
[7] Jagielski, R. and Gero, J. S. (1997). A genetic programming approach to the space layout planning problem.
CAAD Futures, 875–884.
[8] Baykan, C. A. and Fox, M. S. (1997). Spatial synthesis by disjunctive constraint satisfaction. Artificial
Intelligence in Engineering Design, 11, 245–262.
[9] Schwarz, A., Berry, D. M. and Shaviv, E. (1994). Representing and solving the automated building design
problem. Computer-Aided Design, 26(9), 689–698.
[10] Medjdoub, B. and Yannou, B. (1999). Separating topology and geometry in space planning. Computer-Aided
Design, 32, 39–61.
[11] Yin, S. and Cagan, J. (2000). An extended pattern search algorithm for three-dimensional component layout.
Transactions of the ASME, 122, 102–108.
[12] Cagan, J., Degentesh, D. and Yin, S. (1998). A simulated annealing-based algorithm using hierarchical models
for general three-dimensional component layout. Computer-Aided Design, 30(10), 781–790.
[13] Szykman, S. and Cagan, J. (1997). Constrained three-dimensional component layout using simulated annealing.
ASME Transactions, 119, 28–35.
[14] Choo, H. J. and Tommelein, I. D. (1999). Space scheduling using flow analysis. Proc. Seventh Annual
Conference of the International Group for Lean Construction, 299–312.
[15] Rong, B. (1997). Automated generation of fixture configuration design. Journal of Manufacturing Science and
Engineering, 119, 208–219.
[16] Koide, T. and Wakabayashi, S. (1999). A timing-driven floorplanning algorithm with the Elmore delay model for
building block layout. Integration, the VLSI Journal, 27, 57–76.
[17] Wang, Y. and Wu, H. (1998). Method of constraint graphs used in spatial layout. Ruan Jian Xue Bao Journal
of \Software, 9(3), 200–205.
[18] Ito, T. (1999). A genetic algorithm approach to piping route path planning. Journal of Intelligent Manufacturing,
10, 103–114.
[19] Fadel, G. M., Sinha, A. and McKee, T. (2001). Packing optimisation using a rubberband analogy. Proceedings of
DETC’01 2001, ASME Design Engineering Technical Conferences, DETC2001=DAC-21051.
[20] Kim, J. J. and Gossard, D. C. (1991). Reasoning on the location of components for assembly packaging. Journal
of Mechanical Design, Transactions of the ASME, 113(4), 402–407.
[21] Chapman, C. D., Saitou, K. and Jakiela, M. J. (1994). Genetic algorithms as an approach to configuration and
topology Design. Journal of Mechanical Design, Transactions of the ASME, 116(4), 1005–1012.
[22] Papalambros, P. Y. and Wilde, D. J. (2000). Principles of Optimal Design–Modeling and Computation, 2nd ed.
Cambridge University Press, Cambridge, England.
[23] Kott, G. J. and Gabriele, G. (1998). A tunnel based method for mixed discrete constrained nonlinear
optimization. ASME Design Automation Conference, DETC98=DAC-5592.
[24] Kott, G. J. (1998). A method for mixed variable constrained nonlinear optimization based on a new function
bounding technique. PhD thesis, Rensselaer Polytechnic Institute.
484 J. J. MICHALEK et al.

[25] Zhou, J. L. and Tits, A. L. (1995). An SQP algorithm for finely discretized continuous minimax problems and
other minimax problems with many objective functions. SIAM Journal of Optimization, 6, 461–487.
[26] Lawrence, C. T., Zhou, J. L. and Tits, A. (1995). User’s guide for CFSQP version 2.3: A C code for solving large
scale constrained nonlinear minimax optimization problem, generating iterates satisfying all inequality
constraints. Institute for Systems Research, University of Maryland, Technical Report TR-94-16r1.
[27] Bartak, R. (1999). Constraint programming: In pursuit of the Holy Grail. http:==kti.ms.mff.cuni.cz=
bartak=constraints.
[28] Wall, M. (1996). GAlib: A C þ þ library of genetic algorithm components-Version 2.4 -Document revision B.
galib-request@mit.edu, http:==lancet.mit.edu=ga=dist=galibdoc.pdf.
[29] Bentley, P. J. (1999). Evolutionary Design by Computers. Morgan Kaufmann Publishers, Inc., San Francisco, CA.
[30] ASHRAE (1997). Fundamentals. American Society of Refrigeration, Heating and Air-conditioning Engineers,
Atlanta, GA.
[31] IESNA (1998). IESNA Lighting Education: Intermediate Level. Illumination Engineers Society of North
America.